THE RANGE OF POSSIBLE ITERATIVE LEARNING CONTROL LAWS TO OPTIMIZE ROBOT TRACKING

Control systems do not precisely execute the command given to them. In most applications the majority of the error is deterministic and repr oducible. Control systems for robots on assembly lines execute the sam e commands many times each day, repeatedly making the same errors. Sta rting in 1984, the field of learning control generates new control met hods that allow the controller to learn from previous experience perfo rming a specific command, in order to have the error in following the command go to zero as the repetitions of the task progress. This paper gives an overview of the learning control methods developed over the last decade by the author and co-workers. These methods are based on c lassical control, adaptive control, numerical optimization, and digita l signal processing concepts. They can all be viewed in the form of an iterative algorithm for finding an input function to a differential e quations which causes the solution to match the desires robot motion a s output - and the methods are designed to converge in spite of poor k nowledge of the equation. It is seen that most of the control theory b ased methods would not come to mind, if one views the problem from the point of view of numerical optimization. Nevertheless, hardware exper iments on a commercial robot show these methods to be very effective, reducing the root mean square of the tracking error in following a lar ge angle high speed maneuver by a factor of 1000 in roughly 10 iterati ons. This is close the the hardware reproducibility level which is a l ower bound on the tracking accuracy obtainable by any method.